Insights
What Is Jev? TypeSafe AI's New "System One" Model
Casey Rowland
TL;DR:
Jev is a new AI model from TypeSafe AI that doesn't generate text. It outputs typed, structured decisions with confidence scores instead, in as little as 70ms.
TypeSafe calls this a "System One model," built for fast, intuitive machine decisions rather than conversation.
On TypeSafe's own benchmark, Jev roughly matches GPT-5.6 Terra on accuracy but trails GPT-5.6 Sol and Claude Opus 5 by 6-7 points, at a fraction of the cost.
The "can't hallucinate" claim is technically true but narrower than it sounds: Jev can still pick the wrong answer, it just can't invent one that isn't already in its schema.
TypeSafe exited stealth on 2026-09-15 with $40M in seed funding led by DCVC, founded by ex-OpenAI researcher Diogo Almeida.
TypeSafe AI, a startup founded by Diogo Almeida (a co-inventor of the RLHF technique behind ChatGPT), came out of two years of stealth on September 15, 2026 with its first model, Jev. It's built to do something most AI models don't bother with: not talk.
What Jev actually does
TypeSafe describes Jev as "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out." In plain terms, you give it messy, real-world input, and instead of writing a sentence back, it returns a structured value, like a category, a score, or a routing decision, along with a calibrated confidence score for that answer.
This is a real architectural change from how other large language models work. An LLM like GPT or Claude generates text one token at a time, predicting the next word based on everything before it. That's flexible but slow, and it's why a chatbot reply takes a few seconds to stream in.
Jev skips that entirely. It uses what TypeSafe calls a "parallel sampler" that evaluates all possible outputs from a predefined schema at once and returns the answer directly, with no text to generate and nothing to parse afterward.
TypeSafe calls this category a "System One model," a reference to psychologist Daniel Kahneman's research on fast, intuitive thinking (System 1) versus slow, deliberate reasoning (System 2). The pitch is that most software doesn't need an AI to write a paragraph, it needs an AI to make a fast, reliable call: is this support ticket urgent, does this transaction look fraudulent, which of these five categories does this item belong to. That's the job Jev is built for.
How it compares to a normal LLM
Jev (System One model) | A typical frontier LLM | |
|---|---|---|
Output | A typed value from a fixed schema (category, score, decision) | Free-form text, generated token by token |
Speed | 70ms-500ms, per TypeSafe | 3-329 seconds, per TypeSafe's own comparison |
Cost | $0.042 per million input tokens, output tokens free | Varies by model, generally far higher per decision |
Output errors | 0% by construction (only valid schema values are possible) | 0.58%-45.5%, depending on the model, per TypeSafe's testing |
Best suited for | Classification, routing, scoring, structured extraction | Open-ended conversation, writing, reasoning through novel problems |
Those speed and error numbers are what TypeSafe's quotes. Weav is still doing our own independent testing, but we have seen improvements within our own system with Jev. vs. a traditional LLM.
What the benchmarks actually show
On TypeSafe's own four-workflow evaluation, Jev scored 67.8% accuracy, just about matching GPT-5.6 Terra's 67.9%, but trailing both GPT-5.6 Sol (74.1%) and Claude Opus 5 (73.1%) by 6-7 percentage points. Where Jev pulls ahead sharply is cost: roughly $0.0004 per decision, against $0.0304 to $0.1761 for the frontier models it was compared against. TypeSafe's own workflow tests claim up to 193.6x faster and 444.6x cheaper results, though the company itself flags that figure as the high end of real-world gains, not a typical result.
That combination, a little behind the best frontier models on raw accuracy but dramatically cheaper and faster, is the actual tradeoff Jev is offering. It's a reasonable one for high-volume, low-complexity decisions. It's a different question entirely for anything that needs real reasoning.
Does it really "never hallucinate"?
TypeSafe's boldest claim is that Jev is mathematically incapable of hallucinating, because its answers are always pulled from a predefined set of valid schema values. There's nothing for it to invent.
That's true, as far as it goes. But independent technical reviewers have pointed out that it's a narrower claim than it sounds. Jev can still pick the wrong answer from its valid options with high confidence, which produces the same practical outcome for a user as a hallucination would: a confidently wrong response. One widely read technical review of the launch called the "no hallucination" framing "a semantic dodge" for exactly this reason. The same review also raised a separate, more structural question: because Jev can't do the step-by-step reasoning that frontier LLMs use on hard problems, its ceiling on genuinely difficult tasks is likely lower, no matter how calibrated its confidence scores are.
There's also an open question about durability. The same reviewer argued that fast structured output isn't entirely new, pointing to existing techniques like constraining a standard LLM to a fixed set of tokens, and suggested Jev's speed advantage may come more from a clever inference strategy than from a fundamental breakthrough other labs can't replicate. That's a real, open debate, not a resolved one, and it's worth watching rather than assuming either direction.
What this means if you're evaluating AI for customer support
Jev isn't built for conversation, and TypeSafe isn't positioning it that way. But the underlying idea, using a fast, cheap, purpose-built model for a narrow decision instead of a full conversational LLM, is directly relevant to how support automation actually works under the hood. A support AI doesn't use one model for everything: routing a ticket, classifying sentiment, and drafting a full reply are different jobs with different requirements, and treating them as the same call is where a lot of AI support tools quietly overspend. See our breakdown of which AI model actually matters for customer support (/blog/best-ai-model-for-customer-support) for how that tradeoff plays out in practice, and what drives AI model cost (/blog/ai-model-cost) for why locking a support tool to a single model, of any kind, usually costs you more than it needs to.
Where Weav fits
We are currently testing Jev to implement behind the scenes, but we also think it's a genuinely interesting release worth understanding if you're paying attention to where AI is headed. What we do agree with is the underlying premise: not every decision in a support conversation needs the same model. Weav already routes each customer question to the model that fits it, so you're not paying frontier prices for a question a smaller, faster model could answer just as well. If that's the kind of AI-cost problem you're trying to solve, see how Weav handles it (https://weav.com/product).
FAQ
What is Jev? Jev is an AI model from TypeSafe AI that outputs typed, structured decisions with confidence scores instead of generating text, built for fast machine-to-machine decisions rather than conversation.
Who made Jev? TypeSafe AI, a startup founded by Diogo Almeida, a former OpenAI researcher and co-inventor of the RLHF technique used to train ChatGPT. The company exited stealth on September 15, 2026 with $40M in seed funding led by DCVC.
Is Jev faster than ChatGPT or Claude? For the narrow task it's built for, typed decisions from a fixed set of options, yes, dramatically so: TypeSafe reports 70ms-500ms response times versus 3-329 seconds for comparable frontier-model tasks. It's not built to hold a conversation or write open-ended text.
Does Jev really never hallucinate? It can't invent an answer outside its predefined schema, which is a real, structural guarantee. But it can still confidently choose the wrong valid answer, which produces the same practical problem for a user as a hallucination. Independent reviewers have flagged the "never hallucinates" framing as narrower than it sounds.
Is Jev available to use? As of September 2026, Jev is in early access behind a waitlist, with TypeSafe bringing developers off it gradually.
Related Articles
Casey Rowland
